Markdown, blocks JSON, words, and Docling JSON output

Markdown, blocks JSON, words, and Docling JSON output Preview

Structured extraction returns more than one output shape. Markdown comes from extract_markdown, blocks from extract_blocks (and to_json in Rust), words with boxes from extract_words, and a Docling-shaped JSON (DoclingDocument) from extract_docling. Chunks for retrieval pipelines come from iter_chunks. Structured extraction is Preview: output quality is still improving and the API may change in a minor release.

from pdfluent import Document
from pdfluent.chunks import iter_chunks

with Document("report.pdf") as doc:
    md = doc.extract_markdown()                        # fast tier, no models
    md_tables = doc.extract_markdown(quality="balanced")  # needs installed models
    words = doc.extract_words()
    for chunk in iter_chunks(doc, mode="markdown"):
        print(chunk.text)

1. Markdown output

   extract_markdown returns Markdown: headings, list items, table cells, running heads. In Python the default quality="fast" runs the engine alone, pure Rust, with no models, and works everywhere including WebAssembly. quality="balanced" runs a layout model and a table-structure model locally through ONNX Runtime and is Markdown only: extract_markdown reads it, extract_blocks refuses it.

2. Blocks JSON output

   extract_blocks returns blocks: paragraphs, headings with level, list items, table cells with grid position, running heads, each with page and box. In Rust, to_json(&blocks) renders the blocks as JSON. In WebAssembly and .NET, extractBlocks returns a JSON string.

3. Words output

   extract_words returns words with boxes, for pipelines that need token-level positions rather than block structure. Python, Node, WebAssembly, Java and .NET all expose it.

4. Docling JSON output

   extract_docling returns a Docling-shaped JSON (DoclingDocument) for pipelines built on Docling. Python, Node, WebAssembly, Java and .NET all expose it; in Python the call takes first_page, page_count and name.

• Structured extraction is Preview: output quality is still improving and the API may change in a minor release. Plain-text extraction is not Preview.
• The balanced tier is Preview and Markdown only. It runs a layout model and a table-structure model locally through ONNX Runtime 1.23.2 [B49, 2026-10-10]; native only, never WebAssembly; glibc Linux, macOS and Windows; on musl Linux loading returns a typed ModelError.
• Models install once, explicitly: pdfluent models install layout tables. The pdfluent command comes from the CLI archives on the 1.1.0 release page; pip install pdfluent==1.1.0 and npm install @pdfluent/node do not install it. Nothing downloads automatically at run time. Without installed models, the balanced tier refuses with MODEL_NOT_AVAILABLE.
• Table structure (A38, A39, A40; measured 2026-10-14 on the tuning split): balanced (Preview) with the table model scores higher than Docling and pymupdf4llm on FinTabNet and PubTables-1M, and higher than pymupdf4llm on ParseBench.
• Office input to Markdown is covered by claims A31–A34 (measured 2026-10-04). In Python, pdfluent.extract_office(path_or_bytes, format="markdown") takes .docx, .pptx and .xlsx, routing by the file's content types rather than the extension; format="docling" returns DoclingDocument JSON.